Translator

Three-mode translation skill: quick for direct translation, normal for analysis-informed translation, refined for full publication-quality workflow with review and polish.

User Input Tools

When this skill prompts the user, follow this tool-selection rule (priority order):

1. Prefer built-in user-input tools exposed by the current agent runtime — e.g., AskUserQuestion, request_user_input, clarify, ask_user, or any equivalent.
2. Fallback: if no such tool exists, emit a numbered plain-text message and ask the user to reply with the chosen number/answer for each question.
3. Batching: if the tool supports multiple questions per call, combine all applicable questions into a single call; if only single-question, ask them one at a time in priority order.

Concrete AskUserQuestion references below are examples — substitute the local equivalent in other runtimes.

Script Directory

Scripts in scripts/ subdirectory. {baseDir} = this SKILL.md's directory path. Resolve ${BUN_X} runtime: if bun installed → bun; if npx available → npx -y bun; else suggest installing bun. Replace {baseDir} and ${BUN_X} with actual values.

Preferences (EXTEND.md)

Check EXTEND.md in priority order — the first one found wins:

EXTEND.md supports: default target language, default mode, target audience, custom glossaries (inline or file path), translation style, chunk settings.

Schema: references/config/extend-schema.md.

First-Time Setup (BLOCKING)

CRITICAL: When EXTEND.md is not found, you MUST run the first-time setup before ANY translation. This is a BLOCKING operation.

Full reference: references/config/first-time-setup.md

Use AskUserQuestion with all questions (target language, mode, audience, style, save location) in ONE call. After user answers, create EXTEND.md at the chosen location, confirm "Preferences saved to [path]", then continue.

Defaults

All configurable values in one place. EXTEND.md overrides these; CLI flags override EXTEND.md.

Modes

Default mode: Normal (can be overridden in EXTEND.md default_mode setting).

Style presets — control the voice and tone of the translation (independent of audience):

Custom style descriptions are also accepted, e.g., --style "poetic and lyrical".

Auto-detection:

• "快翻", "quick", "直接翻译" → quick mode
• "精翻", "refined", "publication quality", "proofread" → refined mode
• Otherwise → default mode (normal)

Upgrade prompt: After normal mode completes, display:

Translation saved. To further review and polish, reply "继续润色" or "refine".

If user responds, continue with review → polish steps (same as refined mode Steps 4-6 in refined-workflow.md) on the existing output.

Audience presets:

Custom audience descriptions are also accepted, e.g., --audience "AI感兴趣的普通读者".

Workflow

Step 1: Load Preferences

1.1 Check EXTEND.md (see Preferences section above)

1.2 Load built-in glossary for the language pair if available:

• EN→ZH: references/glossary-en-zh.md

1.3 Merge glossaries: EXTEND.md glossary (inline) + EXTEND.md glossary_files (external files, paths relative to EXTEND.md location) + built-in glossary + --glossary file (CLI overrides all)

Step 2: Materialize Source & Create Output Directory

Materialize source (file as-is, inline text/URL → save to translate/{slug}.md), then create output directory: {source-dir}/{source-basename}-{target-lang}/. Detect source language if --from not specified.

Full details: references/workflow-mechanics.md

Output directory contents (all intermediate and final files go here):

Step 3: Assess Content Length

Quick mode does not chunk — translate directly regardless of length. Before translating, estimate word count. If content exceeds chunk threshold (default 4000 words), proactively warn: "This article is ~{N} words. Quick mode translates in one pass without chunking — for long content, --mode normal produces better results with terminology consistency." Then proceed if user doesn't switch.

For normal and refined modes:

3.1 Long Content Preparation (normal/refined modes, >= chunk threshold only)

Before translating chunks:

1. Extract terminology: Scan entire document for proper nouns, technical terms, recurring phrases
2. Build session glossary: Merge extracted terms with loaded glossaries, establish consistent translations
3. Split into chunks: Use ${BUN_X} {baseDir}/scripts/main.ts <file> [--max-words <chunk_max_words>] [--output-dir <output-dir>]
   • Parses markdown blocks (headings, paragraphs, lists, code blocks, tables, etc.)
   • Splits at markdown block boundaries to preserve structure
   • If a single block exceeds the threshold, falls back to line splitting, then word splitting
4. Assemble translation prompt:
   • Main agent reads 01-analysis.md (if exists) and assembles shared context using Part 1 of references/subagent-prompt-template.md — inlining: target style, content background, merged glossary, and translation challenges
   • Save as 02-prompt.md in the output directory (shared context only, no task instructions)
5. Draft translation via subagents (if Agent tool available):
   • Spawn one subagent per chunk, all in parallel (Part 2 of the template)
   • Each subagent reads 02-prompt.md for shared context, receives chunk position info (chunk N of M + brief context of where it sits in the argument), translates its chunk, saves to chunks/chunk-NN-draft.md
   • Consistency is guaranteed by the shared 02-prompt.md (glossary, figurative language mapping, comprehension challenges, source voice, and translation challenges from analysis)
   • If no chunks (content under threshold): spawn one subagent for the entire source file
   • If Agent tool is unavailable, translate chunks sequentially inline using 02-prompt.md
6. Merge: Once all subagents complete, combine translated chunks in order. If chunks/frontmatter.md exists, prepend it. Save as 03-draft.md (refined) or translation.md (normal)
7. All intermediate files (source chunks + translated chunks) are preserved in chunks/

After chunked draft is merged, return control to main agent for critical review, revision, and polish (Step 4).

Step 4: Translate & Refine

Translation principles (apply to all modes):

• Rewrite, not translate: Rewrite content into natural, engaging target language as if a skilled native writer composed it from scratch. Quality test: "Does this read like it was originally written in the target language?"
• Accuracy first: Facts, data, and logic must match the original exactly
• Natural flow: Use idiomatic target language word order. Break long source sentences into shorter, natural ones. Interpret metaphors and idioms by intended meaning, not word-for-word
• Terminology: Use standard translations consistently. First occurrence of specialized terms: annotate with original in parentheses
• Preserve format: Keep all markdown formatting (headings, bold, italic, images, links, code blocks)
• Proactive interpretation: For jargon or concepts the target audience may lack context for, add concise explanations in bold parentheses （**解释**）. Keep annotations few — only where genuinely needed for comprehension
• Frontmatter: If source has YAML frontmatter, rename source-metadata fields with source prefix (camelCase: url→sourceUrl, title→sourceTitle, etc.), add translated values as new top-level fields (skip title if body has H1), keep other fields as-is

Quick Mode

Translate directly → save to translation.md. Apply all translation principles above.

Normal Mode

1. Analyze → 01-analysis.md (domain, tone, terminology, translation challenges)
2. Assemble prompt → 02-prompt.md (translation instructions with context, glossary, challenges)
3. Translate (following 02-prompt.md) → translation.md

After completion, prompt user: "Translation saved. To further review and polish, reply 继续润色 or refine."

If user continues, proceed with critical review → revision → polish (same as refined mode Steps 4-6 below), saving 03-draft.md (rename current translation.md), 04-critique.md, 05-revision.md, and updated translation.md.

Refined Mode

Full workflow for publication quality. See references/refined-workflow.md for detailed guidelines per step.

The subagent (if used in Step 3.1) only handles the initial draft. All subsequent steps (critical review, revision, polish) are handled by the main agent, which may delegate to subagents at its discretion.

Steps and saved files (all in output directory):

1. Analyze → 01-analysis.md (domain, tone, terminology, translation challenges)
2. Assemble prompt → 02-prompt.md (translation instructions with inlined context)
3. Draft → 03-draft.md (initial translation with translator's notes; from subagent if chunked)
4. Critical review → 04-critique.md (diagnosis only: accuracy, Europeanized language, strategy execution, expression issues)
5. Revision → 05-revision.md (apply all critique findings to produce revised translation)
6. Polish → translation.md (final publication-quality translation)

Each step reads the previous step's file and builds on it.

Step 5: Output

Final translation is always at translation.md in the output directory.

After the final translation is written, do a lightweight image-language pass:

1. Collect image references from the translated article
2. Identify likely text-heavy images such as covers, screenshots, diagrams, charts, frameworks, and infographics
3. If any image likely contains a main text language that does not match the translated article language, proactively remind the user
4. The reminder must be a list only. Do not automatically localize those images unless the user asks

Reminder format (use whatever image syntax the article already uses — standard markdown or wikilink):

Possible image localization needed:
- ![example cover](attachments/example-cover.png): likely still contains source-language text while the article is now in target language
- ![example diagram](attachments/example-diagram.png): likely text-heavy framework graphic, check whether labels need translation

Display summary:

**Translation complete** ({mode} mode)

Source: {source-path}
Languages: {from} → {to}
Output dir: {output-dir}/
Final: {output-dir}/translation.md
Glossary terms applied: {count}

If mismatched image-language candidates were found, append a short note after the summary telling the user that some embedded images may still need image-text localization, followed by the candidate list.

Extension Support

Custom configurations via EXTEND.md. See Preferences section for paths and supported options.